Editorial Policy

This page explains how Local Model Watch selects, writes, publishes and corrects its articles, so that readers can see how the site’s content is produced.

Scope

We cover open-weight generative models (text, image, video and audio) that can be run on your own PC or server, together with the inference engines, quantization and fine-tuning tools used to run them. Models that are not meant for generation, such as embedding models and classifiers, are out of scope. See About this site for details.

Sources and selection criteria

  • Our sources are official announcements from major inference-engine projects and
    model publishers (Hugging Face, GitHub, official blogs), and discussions in developer communities such as Hacker News and Bluesky.
  • Candidates are filtered by mechanical criteria: whether the publisher is on a
    curated list of notable organizations and developers, or whether downloads and engagement are rising sharply.
  • Every article lists the URLs it was written from under “Sources".
    An article with no source URL is never published.
  • Facts and figures are limited to what the sources say. We do not add anything the
    sources do not contain; unknown items are omitted rather than marked “unknown".

What we do not publish

  • Topics whose sources do not support an article of reasonable length. A program
    measures the finished text and withholds any article below a per-type minimum.
  • Patch releases of inference engines (small bug-fix versions) do not get their own
    article unless the release notes are substantial. They appear in the engines and tools release tracker and in the “Releases Since Our Last Article" table of the next article.
  • Hacker News and Bluesky items are covered only when they mention something that
    signals running models locally — a model name, an inference engine, quantization and the like. General AI discussion is out of scope.
  • Text models published by individuals not on our notable list are not given their own
    article even when trending; they are listed in the weekly roundup instead.

“Verified" and “Unverified"

Each article carries a reliability tag.

  • Verified: confirmed against a primary source (the official repository or announcement).
  • Unverified: based on community posts or other secondary sources and not yet
    confirmed. A notice appears at the top of the article. Once confirmed, we append an update note and remove the notice.

Figures are fixed by code, not by AI

Generative AI can invent plausible-looking numbers. For that reason, the figures that matter most in our articles are computed and fixed by a program, separately from the writing step.

Memory requirements and hardware

The “Hardware Requirements" table in each model article is produced as follows:

  1. Take the actual sizes of the files distributed on Hugging Face, summed per
    quantization type (split files are added together).
  2. Add a 20% runtime overhead for the KV cache and similar allocations.
  3. Determine which common GPU memory tier (8GB / 12GB / 16GB / 24GB / 32GB / 48GB /
    80GB) the build fits in, and show the largest build that fits in each tier.

If no distributed files are available and only the parameter count is known, we estimate from the full-precision size using the same steps. These are our own estimates, not figures quoted from the model card; actual usage varies with context length, batch size and inference engine. To compare across models, see our Local Models by VRAM quick reference.

Units and benchmarks

  • Large numbers (“billion", “million") are converted into Japanese units by code, not by
    the writer.
  • We explain what a benchmark measures only when it appears in our curated glossary.
    For benchmarks not in the glossary we report score comparisons without interpretation.
  • Table figures are copied verbatim from the sources, without rounding or conversion.

Quantized and converted variants

When a format conversion (GGUF, MLX, AWQ, etc. — same weights, different format or precision) of a model we have already covered appears later, we do not publish a separate article. Instead we append it to the original article’s “Quantized and Converted Variants" section, because the only new information — who converted it, into what format, and which GPU it runs on — can be determined mechanically as described above. Fine-tunes and other models with changed weights are covered as separate models.

Related articles and weekly roundups

  • The “Related Articles" list at the end of each article is selected mechanically from
    our own records: articles about the same model family, the same publisher, the same tools, or the same VRAM tier.
  • “This Week in Numbers" at the top of each weekly roundup is computed by a program from
    the records of the articles published that week.
  • “Recent Models in the Same Size Class" in new-model articles lists our earlier articles
    in the same parameter range, selected mechanically from our records (the same family is handled by “Related Articles" instead).
  • The model family pages, the engines and tools tracker
    and the glossaries are likewise assembled by programs from our article records, GitHub release lists and definitions kept in the site’s source code. The one-line project descriptions and glossary definitions are written by hand.

Use of AI and the role of the editor

Local Model Watch uses AI agents to collect and organize information, draft articles and translate them. The site’s operator (an individual) designs and maintains the sources, selection criteria, article structure, calculation methods and correction policy, and reviews published articles. The parts that articles depend on — the presence of sources, the reliability tag, computed figures and whether an image may be shown — are enforced by programmatic checks rather than left to the AI’s judgment.

When we get something wrong

If an article turns out to be wrong, we do not delete it. We add a correction note at the top so that readers can see what was corrected and how. Please report errors through the contact page.

Advertising

This site may display advertising (such as Google AdSense). Advertising does not influence what we cover or how we cover it. See our privacy policy for details.